Mechanical cross-domain fault diagnosis method based on parameter progressive integration
By employing a parameter-incremental ensemble method with multi-scale stylization enhancement and mutual information constraints, the problems of feature distortion and performance fluctuation in cross-domain mechanical fault diagnosis are solved, thereby improving the accuracy and robustness of mechanical fault diagnosis.
Patent Information
- Application Number
- CN202511778536.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-02-27
AI Technical Summary
Existing technologies for cross-domain fault diagnosis in machinery suffer from severe feature distortion, large fluctuations in model training performance, and poor cross-domain generalization ability. In particular, in scenarios where there is a lack of target domain data and insufficient diversity of source domain data, it is difficult to effectively simulate the distribution of the target domain and maintain feature consistency.
A multi-scale stylization enhancement module is used to generate pseudo-domain signals, construct task models and surrogate models, maintain feature consistency through mutual information constraints, and accumulate cross-domain knowledge through parameter progressive integration to reduce training performance fluctuations.
It significantly improves the accuracy and robustness of mechanical fault diagnosis, effectively simulates the target domain distribution, maintains feature semantic consistency, and reduces the instability of model training.
Smart Images
Figure CN121579984A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to a mechanical cross-domain fault diagnosis method based on parameter progressive integration, belonging to the technical field of intelligent diagnosis of mechanical equipment and deep learning cross-domain migration. BACKGROUND
[0002] Mechanical equipment is the core of modern industrial production, and its running state is directly related to production safety and economic benefits. Therefore, accurate and timely fault diagnosis is crucial. Traditional mechanical fault diagnosis methods mainly rely on manual feature extraction combined with shallow machine learning models, which can achieve certain results under specific working conditions, but have limited generalization ability and are difficult to adapt to complex and variable industrial environments. With the development of deep learning technology, fault diagnosis methods based on deep neural networks have made significant progress due to their powerful automatic feature extraction capabilities. However, most deep learning models assume that training data and test data follow independent and identically distributed. In practical applications, due to changes in operating conditions (such as speed, load), environmental factor fluctuations, equipment aging, and differences between different devices, there are significant differences in data distribution between mechanical vibration signals collected under different working conditions, i.e., domain shift problem. This makes the model trained under a certain working condition (source domain) have a sharp decline in diagnostic performance when applied to unseen working conditions (target domain). To solve the domain shift problem, domain generalization methods have emerged, which aim to train models using only source domain data, allowing them to directly generalize to unseen target domains. In particular, in the single-source domain generalization scenario where there is a lack of target domain data and insufficient diversity of source domain data, how to achieve effective cross-domain fault diagnosis has become a major challenge in the field of intelligent diagnosis.
[0003] Currently, data augmentation strategy is the mainstream solution to solve the single-source domain generalization problem. Data augmentation generates pseudo-domain samples with different distribution characteristics by transforming source domain data to simulate the distribution offset of the target domain. However, existing data augmentation methods have significant limitations: first, simple augmentation strategies are difficult to effectively simulate the complex domain offset patterns in real industrial scenarios; second, complex augmentation operations may introduce noise or cause feature distortion, i.e., the semantic consistency between the augmented samples and the original samples is lost, making the feature representation learned by the model not robust enough. More seriously, domain generalization models based on data augmentation generally have performance fluctuations during training, i.e., the model's diagnostic performance on the target domain fluctuates at different stages of training. In-depth analysis shows that this is mainly due to: the model's difficulty in effectively accumulating and integrating knowledge learned from diverse augmented data; the feature representation obtained in previous training steps is easily covered or distorted by subsequent augmented samples; the lack of effective mechanisms to ensure the consistency of the feature space when learning different augmented samples. This training instability greatly reduces the reliability and robustness of the model. In terms of model training strategies, some methods attempt to improve generalization by parameter integration, such as exponential moving average, random weight average, etc. However, traditional parameter averaging methods lack effective regularization constraints, making it difficult to ensure the consistency of feature representation between models at different training stages. In particular, in the dual model collaborative training architecture, how to effectively constrain the consistency of feature representation between different models, and avoid model divergence or loss of semantic information, is still a key problem to be solved. In summary, existing methods still face core challenges when dealing with mechanical cross-domain fault diagnosis, especially in scenarios where there is a lack of target domain data and the diversity of source domain data is insufficient: how to design an effective data augmentation strategy that can simulate the target domain distribution while avoiding feature distortion; how to stably accumulate cross-domain knowledge during training to reduce performance fluctuations; how to maintain feature representation consistency in a dual model architecture; how to achieve gradual accumulation of knowledge through parameter space integration.
[0004] Therefore, there is an urgent need in the field for a new mechanical cross-domain fault diagnosis method that can generate high-quality pseudo-domain samples through multi-scale style enhancement, alleviate feature distortion through a dual-model architecture combined with mutual information constraints, and achieve gradual accumulation of cross-domain knowledge through parameter space integration, significantly reducing training performance fluctuations and improving the accuracy and robustness of cross-domain diagnosis. SUMMARY
[0005] The present application aims to provide a mechanical cross-domain fault diagnosis method based on parameter progressive integration, which aims to solve the technical problems of severe feature distortion, large model training performance fluctuations and poor cross-domain generalization ability in mechanical cross-domain fault diagnosis under the condition of lack of target domain data and insufficient diversity of source domain data, significantly improving the accuracy and robustness of mechanical fault diagnosis.
[0006] To achieve the above-mentioned purpose, the technical solution of the present application is: a mechanical cross-domain fault diagnosis method based on parameter progressive integration, comprising the following steps: Step 1: input the source domain mechanical vibration signal into a multi-scale stylization enhancement module to generate a pseudo-domain mechanical vibration signal; wherein the multi-scale stylization enhancement module includes multi-scale convolution operation, adaptive instance normalization operation and multi-scale deconvolution operation; Step 2: construct a task model and a proxy model; wherein the task model and the proxy model are both composed of a feature extractor and a classifier, the feature extractor adopts a neural network structure composed of multiple one-dimensional convolution, pooling and nonlinear activation unit, and the classifier is composed of several fully connected layers and a Softmax unit; Step 3: input the source domain mechanical vibration signal into the task model to obtain a source domain feature vector, input the pseudo-domain mechanical vibration signal into the proxy model to obtain a pseudo-domain feature vector, construct a contrastive loss function based on the source domain feature vector and the pseudo-domain feature vector, and use the contrastive loss function to impose mutual information constraints on the task model and the proxy model; Step 4: after completing the mutual information constraints, periodically update the parameter trajectory of the proxy model to the task model after averaging, obtain the final task model, and diagnose the mechanical fault of the unseen target domain based on the final task model.
[0007] Optionally, the Step 1 is specifically: The source domain mechanical vibration signal is defined as , wherein represents the th source sample, is its corresponding class label, and there are a total of samples; A multi-scale convolution operation is performed, and a multi-scale convolution layer is adopted Local features and global features of different time scales are extracted from the source domain mechanical vibration signal to obtain multi-scale features. An adaptive instance normalization operation is performed, and the mean and variance of the multi-scale features are adjusted by injecting additive noise and multiplicative noise to obtain style-enhanced multi-scale features. A multi-scale deconvolution operation is performed, and a multi-scale deconvolution layer corresponding to the multi-scale convolution layer is adopted The style-enhanced multi-scale features are restored to multi-scale style-enhanced mechanical vibration signals of the same length as the source domain mechanical vibration signal. The source domain mechanical vibration signal and the multi-scale style-enhanced mechanical vibration signal are weighted and mixed, and after Sigmoid function normalization processing, a pseudo-domain mechanical vibration signal is obtained .
[0008] Optionally, the multi-scale convolution layer adopts convolution kernels of lengths 3, 5, 7, and 9.
[0009] Optionally, a contrastive loss function is constructed based on the source domain feature vector and the pseudo-domain feature vector, and mutual information constraints are imposed on the task model and the agent model using the contrastive loss function, specifically: The source domain feature vector and the pseudo-domain feature vector are combined to form a real paired sample , . The order of all pseudo-domain feature vectors in a batch is randomly shuffled to obtain an error paired feature , a pseudo-paired sample , is constructed, and a negative sample for mutual information estimation is obtained. Based on the negative sample, a contrastive loss function is constructed, and mutual information constraints are imposed on the task model and the agent model by maximizing the mutual information of the real paired sample and minimizing the mutual information of the pseudo-paired sample.
[0010] Optionally, the expression of the contrastive loss function is: wherein, is the contrastive loss, is the similarity.
[0011] Optionally, obtaining the final task model specifically comprises: The initial parameters of the task model are copied to obtain the initial parameters of the agent model; and parameter snapshots of the agent model are continuously recorded To form a parameter trajectory, every k training rounds, the arithmetic mean of all parameters in the parameter trajectory is calculated element by element to obtain the average parameter. The average parameter is then used to update the task model parameters. After iterating for a preset number of rounds, the final task model is obtained.
[0012] The beneficial effects of this invention are: 1. By using multi-scale stylization to enhance the generation of diverse pseudo-domain mechanical vibration signal samples, the working condition changes in the target domain can be effectively simulated. Under the condition of lack of target domain data and insufficient diversity of source domain data, the cross-domain generalization ability of the model is significantly enhanced.
[0013] 2. By combining a dual-model architecture with a mutual information constraint mechanism, the semantic consistency between the original features and the augmented features is maintained, which effectively alleviates the feature distortion problem caused by data augmentation and enhances the robustness of the model features.
[0014] 3. By periodically averaging and updating model parameters through progressive parameter integration, cross-domain knowledge is gradually accumulated, which significantly reduces performance fluctuations during model training and improves the stability and reliability of the diagnostic model. Attached Figure Description
[0015] Figure 1 This is a flowchart of the steps of the present invention; Figure 2 This is a detailed flowchart of the multi-scale style enhancement and parametric progressive integration of the present invention. Detailed Implementation
[0016] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0017] Example 1: As Figure 1 As shown, a cross-domain fault diagnosis method for machinery based on parameter progressive integration includes the following steps: Step 1: Input the source domain mechanical vibration signal into the multi-scale stylization enhancement module to generate a pseudo-domain mechanical vibration signal; wherein, the multi-scale stylization enhancement module includes multi-scale convolution operation, adaptive instance normalization operation and multi-scale deconvolution operation; Optionally, the source domain mechanical vibration signal is defined as ,in, Indicates the first One source sample, For its corresponding category labels, a total of One sample; Perform multi-scale convolution operations and use multi-scale convolutional layers. Local features and global features of different time scales are extracted from the source domain mechanical vibration signal to obtain features of different scales, and multi-scale features are obtained; specifically, in this embodiment, the multi-scale convolution layer uses convolution kernels with lengths of 3, 5, 7, and 9; An adaptive instance normalization operation is performed to adjust the mean and variance of the multi-scale features by injecting additive noise and multiplicative noise, so that the original semantics are maintained while the style of the controllable source domain mechanical vibration signal is enhanced, the domain distribution change under different working conditions is simulated, and the multi-scale feature with enhanced style is obtained; A multi-scale deconvolution operation is performed through a multi-scale deconvolution layer corresponding to the multi-scale convolution layer The multi-scale feature with enhanced style is restored to a multi-scale style-enhanced mechanical vibration signal with the same length as the source domain mechanical vibration signal; The source domain mechanical vibration signal and the multi-scale style-enhanced mechanical vibration signal are weighted and mixed, and after Sigmoid function normalization processing, a pseudo-domain mechanical vibration signal is obtained to simulate the distribution offset of the target domain.
[0018] Step2: Construct a task model and a proxy model; wherein the task model and the proxy model both consist of a feature extractor and a classifier, the feature extractor adopts a neural network structure consisting of multiple one-dimensional convolution, pooling and nonlinear activation units, and the classifier consists of several fully connected layers and a Softmax unit; Optionally, the task model consists of a feature extractor and a classifier ; the proxy model consists of a feature extractor and a classifier , and the feature extractor and the classifier can both be composed of known modules.
[0019] Step3: input the source domain mechanical vibration signal into the task model to obtain a source domain feature vector, input the pseudo-domain mechanical vibration signal into the proxy model to obtain a pseudo-domain feature vector, construct a contrast loss function based on the source domain feature vector and the pseudo-domain feature vector, and use the contrast loss function to impose mutual information constraints on the task model and the proxy model; Optionally, the contrast loss function is constructed based on the source domain feature vector and the pseudo-domain feature vector, and the mutual information constraints are imposed on the task model and the proxy model by using the contrast loss function, which is specifically: the source domain feature vector and the pseudo-domain feature vector are used to form a real paired sample , ; randomly shuffle the order of all pseudo-domain feature vectors in a batch to obtain error-paired features , construct pseudo-paired samples , , thereby obtaining negative samples for mutual information estimation; Based on the negative samples, a contrastive loss function is constructed, and by maximizing the mutual information of real pairs and minimizing the mutual information of pseudo pairs, mutual information constraints are imposed on the task model and the agent model.
[0020] Optionally, the expression of the contrastive loss function is: wherein, is the contrastive loss, is the similarity.
[0021] It can be understood that the embodiment realizes consistent feature semantic representation that can be extracted by the task model and the agent model under different data distributions through mutual information constraints, effectively alleviating feature distortion caused by style enhancement.
[0022] As shown in Figure 2 , it is a detailed flowchart of the multi-scale style enhancement and parameter progressive integration of the application, which specifically includes the above Step1-Step3.
[0023] Step4: After completing the mutual information constraint, periodically update the parameter trajectory of the agent model to the task model after averaging, to obtain the final task model, and diagnose mechanical faults of unseen target domains based on the final task model.
[0024] Optionally, the final task model is obtained by: copying the initial parameters of the task model to obtain the initial parameters of the agent model; continuously recording parameter snapshots of the agent model to form a parameter trajectory, and every k training rounds, the arithmetic mean of all parameters in the parameter trajectory is calculated to obtain an average parameter, and the average parameter is updated to the task model parameters, to realize progressive accumulation of cross-domain knowledge and improvement of parameter stability, and after iterating for a preset number of times, the final task model is obtained.
[0025] Optionally, the diagnosis of mechanical faults of unseen target domains based on the final task model is specifically: extracting mechanical vibration signals through the feature extractor of the final task model to obtain a feature vector, and then inputting the feature vector into the classifier of the final task model to output a fault category classification result, thereby realizing cross-domain fault diagnosis of mechanical equipment under different working conditions.
[0026] To further verify the effectiveness of the present application, the present application is compared with other fault diagnosis methods. Accuracy (ACC) is selected to measure the performance of fault diagnosis. Acc
[0027] Specifically, ACC is a commonly used index for classification tasks, and in recent years, it has also been widely used in many fault diagnosis works. Given a set of samples , and the true label of each sample is . ACC index can be calculated by the following formula, the core of which is to count the proportion of samples that are correctly classified among all samples: , wherein is the total number of samples, represents the predicted label of the model for the i-th sample, represents the true label of the i-th sample is an indicator function, the function value is 1 when the condition in the parentheses (i.e. prediction equals true value) is true, otherwise 0. The larger the value of ACC index represents the better the fault diagnosis effect. 。
[0028] Further, the present application selects the public dataset JNU bearing dataset of Jiangnan University for experiment. The experimental platform of Jiangnan University collects vibration data under three different rotating speeds (600, 800 and 1000 rpm). Under each rotating speed, the measurement data of four bearing states are recorded: normal operation and three fault types (inner ring defect, outer ring defect and rolling element damage). The vibration acceleration signal is sampled at a frequency of 5k HZ, and each state contains 400 samples, each sample contains 2048 data points.
[0029] Further, the present experiment sets three main tasks and six sub-tasks for the bearing dataset of Jiangnan University, respectively, the main tasks are denoted as T1, T2, T3, and the sub-tasks are denoted as T1-C2, T1-T3, T2-C1, T2-C3, T3-C1, T3-C2, wherein the main task T1 means that the source domain C1 data is used for training, and the target domain C2, C3 data is tested respectively, and the average of the classification accuracy of C2, C3 is taken as the final result, and the main tasks T2, T3 are the same as above. The sub-task T1-C2 means that in the T1 task, the source domain C1 data is used for training, and only the C2 target domain data is used for testing to obtain the final classification accuracy result, and the other sub-tasks are the same as above. As shown in Table 1, the fault diagnosis task settings for the bearing dataset of Jiangnan University are shown, and the present embodiment repeats the experiment for each of the main task and the sub-task for 5 times, and the experimental results are shown in Table 2.
[0030] Table 1 Bearing dataset of Jiangnan University for fault diagnosis task setting
[0031] Table 2 Fault diagnosis effects of different fault diagnosis methods on the bearing dataset of Jiangnan University
[0032] As can be seen from Table 2, the method of the present application performs best in classification accuracy, indicating that the fault diagnosis effect of the present application is good.
[0033] In summary, the present application first enhances the source domain mechanical vibration signal by using a multi-scale style enhancement module, generates a pseudo domain mechanical vibration signal through adaptive instance normalization and multi-scale convolution transformation, and simulates the distribution offset of the target domain. Secondly, a double model architecture of task model and proxy model is constructed, the task model is learned on the source domain mechanical vibration signal data, the proxy model is learned on the pseudo domain mechanical vibration signal data, and the consistency with the feature representation of the task model is maintained through mutual information constraint, effectively alleviating the feature distortion caused by enhancement. Thirdly, the parameter trajectory of the proxy model is continuously recorded and averaged during the training process, and the average parameters are periodically updated to the task model, realizing the gradual accumulation of cross-domain knowledge. Finally, the updated task model is used to diagnose the mechanical fault of the unseen target domain. The parameter progressive integration method used in the present application has significant advantages in handling mechanical cross-domain fault diagnosis, especially under the condition of lacking target domain data and insufficient diversity of source domain data, the present application not only effectively preserves the semantic consistency of the features through multi-scale style enhancement and mutual information constraint, but also improves the stability of the model during the training process through parameter progressive integration, significantly reduces the performance fluctuation of the model training, and finally the diagnosis result accurately reflects the fault characteristics under different working conditions, further improving the accuracy and robustness of mechanical cross-domain fault diagnosis.
[0034] The specific embodiments of the present application are described in detail above in combination with the drawings, but the present application is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the present application.
Claims
1. A mechanical cross-domain fault diagnosis method based on parameter progressive integration, characterized in that, The method includes the following steps: Step 1: Input the source domain mechanical vibration signal into the multi-scale stylization enhancement module to generate a pseudo-domain mechanical vibration signal; wherein, the multi-scale stylization enhancement module includes multi-scale convolution operation, adaptive instance normalization operation and multi-scale deconvolution operation; Step 2: Construct the task model and the agent model; wherein, the task model and the agent model are both composed of a feature extractor and a classifier. The feature extractor adopts a neural network structure composed of multiple one-dimensional convolution, pooling and non-linear activation units, and the classifier is composed of several fully connected layers and Softmax units. Step 3: Input the source domain mechanical vibration signal into the task model to obtain the source domain feature vector, input the pseudo domain mechanical vibration signal into the surrogate model to obtain the pseudo domain feature vector, construct a contrastive loss function based on the source domain feature vector and the pseudo domain feature vector, and apply mutual information constraints to the task model and the surrogate model using the contrastive loss function; Step 4: After completing the mutual information constraint, the parameter trajectory of the proxy model is periodically averaged and then updated to the task model to obtain the final task model, so as to diagnose mechanical faults in the unseen target domain based on the final task model.
2. The mechanical cross-domain fault diagnosis method based on parameter progressive integration according to claim 1, characterized in that, Step 1 specifically refers to: Define the source domain mechanical vibration signal as ,in, Indicates the first One source sample, For its corresponding category labels, a total of One sample; Perform multi-scale convolution operations and use multi-scale convolutional layers. Local and global features at different time scales are extracted from the source domain mechanical vibration signal to obtain multi-scale features; An adaptive instance normalization operation is performed, and the mean and variance of the multi-scale features are adjusted by injecting additive and multiplicative noise to obtain style-enhanced multi-scale features. Perform multi-scale deconvolution operations, through the multi-scale deconvolution layers corresponding to the multi-scale convolution layers. The style-enhanced multi-scale features are restored to a multi-scale style-enhanced mechanical vibration signal of the same length as the source domain mechanical vibration signal; The source domain mechanical vibration signal and the multi-scale style-enhanced mechanical vibration signal are weighted and mixed, and then normalized using the Sigmoid function to obtain the pseudo-domain mechanical vibration signal. .
3. The mechanical cross-domain fault diagnosis method based on parameter progressive integration according to claim 2, characterized in that, The multi-scale convolutional layers use convolutional kernels of lengths of 3, 5, 7, and 9.
4. The mechanical cross-domain fault diagnosis method based on parameter progressive integration according to claim 3, characterized in that, The specific steps involve constructing a contrastive loss function based on the source domain feature vector and the pseudo-domain feature vector, and applying mutual information constraints to the task model and the proxy model using the contrastive loss function. The source domain feature vector and the pseudo-domain feature vector Composition of real paired samples ( , ); All pseudo-domain feature vectors within the batch The order is randomly shuffled to obtain mismatched features. Construct pseudo-paired samples ( , This allows us to obtain negative samples for mutual information estimation. A contrastive loss function is constructed based on the negative samples. By maximizing the mutual information of real pairings and minimizing the mutual information of pseudo pairings, mutual information constraints are imposed on the task model and the agent model.
5. The mechanical cross-domain fault diagnosis method based on parameter progressive integration according to claim 4, characterized in that, The expression for the contrastive loss function is: ； in, To compare the losses, For similarity.
6. The mechanical cross-domain fault diagnosis method based on parameter progressive integration according to claim 1, characterized in that, The final task model obtained is specifically as follows: The initial parameters of the surrogate model are obtained by copying the initial parameters of the task model; the parameter snapshots of the surrogate model are continuously recorded. To form a parameter trajectory, every k training rounds, the arithmetic mean of all parameters in the parameter trajectory is calculated element by element to obtain the average parameter. The average parameter is then used to update the task model parameters. After iterating for a preset number of rounds, the final task model is obtained.